Adaptive stress testing (AST) package. Finds and analyzes likeliest failures for a simulated system under test. See documentation for usage.

clean()

Forcibly remove temporary environment, purging all enabled submodules. Only necessary if submodule manager is corrupted and disable cannot restore functionality. Takes effect after Julia restart.

disable(submodule)

Disable submodule(s). Accepts string or vector of strings. With zero arguments defaults to all enabled submodules. Takes effect after Julia restart.

enable(submodule)

Enable submodule(s). Accepts string or vector of strings. With zero arguments defaults to all associated submodules. Takes effect immediately.

enabled()

List enabled submodules.

exclude(dir::String)

Associate package directory to main package via submodule table instead of direct code loading. Equivalent of include for submodules / optional dependencies.

Invoke any submodule-related initializations. Called by top-level __init__ function.

load()

Load enabled submodules (necessary after Julia restart). Takes effect immediately.

submodules()

List all available submodules.

List unregistered dependencies of submodule.

Suppresses output if verbosity condition is not met.

Interface between custom simulation and AST framework. Supports various levels of observability and access.

ASTClient

Solver-side client. Interacts with server via TCP.

Standard MDP object for AST. Wraps simulation and contains auxiliary information and parameters.

ASTMDP(sim::AbstractSimulation; kwargs...)

Constructor for ASTMDP object. Infers various properties of MDP. Allows kwargs to be specified flatly and automatically assigns to correct level.

ASTServer

Simulation-side server. Interacts with client via TCP.

ASTServer(mdp::ASTMDP; kwargs...)

Constructor for server.

Exception type for server-side error. Allows exceptions to propogate to client to avoid hanging connection.

Abstract AST MDP type.

AST core objective function abstract type. Defines fundamental reward function by combining log probability, event bonus, and distance heuristic. Additional contributions to the reward should be added by implementing reward function in GrayBox or BlackBox interface.

Distance heuristic abstract type. Custom heuristics inherit this type and implement reset! and apply! functions. Application can return scalar value or partially-applied function transforming f(s,s') to f(s)(s').

Abstract AST reward structure.

Internal abstract type.

Abstract type for AST action.

BlackBox <: AbstractSimulation

Parent type for user simulation in which environment is not exposed.

Set of environment variables and their distributions.

Properties of environment variables.

Samples from environment.

Final distance encountered. If MDP is episodic, this is equal to the minimum distance (determined retroactively instead of accumulated as in MinimumHeuristic).

Environment that will be flattened upon sampling.

Conditional function. Applies function if applicable, otherwise returns value. Useful for conditional partial function application.

Gradient of conservative potential. Default and recommended.

GrayBox <: AbstractSimulation

Parent type for user simulation in which environment is exposed.

Minimum distance across episode. Warning: non-Markovian.

Null heuristic. Returns zero.

State type for observable simulation.

RemoteASTMDP{S<:State, A<:Action} <: AbstractASTMDP{S, A}

MDP object that virtually references remote MDP. Autogenerated by ASTClient.

Standard AST reward structure.

Action type corresponding to instantiation of stochastic environment.

Action type corresponding to setting random seed of simulation.

Abstract type for AST state.

Optional private token for additional security. Use of token prompts seed rehashing to obstruct client-side interpretability.

Internal unimplemented exception type.

State type for unobservable simulation.

Inferred properties of single environment variable.

Vector reward function. Maintains separate components to facilitate post-analysis and enhanced learning methods.

Default objective function. Sums components at each timestep.

Respond to echo request from ASTClient.

Respond to information request from ASTClient.

Apply raw action to environment and return reward.

Apply converted action to environment and return reward.

Infer dimension of action space.

Return action type.

Return set of available actions as sampleable object. # TODO: not type-stable.

Perform function call on server MDP. Call sends request to server with provided function and arguments and receives return value. Blocks until complete.

Close SSH tunnel to/from remote server.

connect!(client::ASTClient; remote::String="", remote_port::Int64=1812)

Connect client to server, optionally through SSH tunnel. Optional keyword argument remote should be of the form user@machine.

connect!(server::ASTServer; remote::String="", remote_port::Int64=1812)

Connect server, optionally through SSH tunnel. Optional keyword argument remote should be of the form user@machine.

Wrap action for internal handling.

Internal handling of custom reward.

disconnect!(client::ASTClient)

Disconnect client from server.

disconnect!(server::ASTServer)

Disconnect server.

distance(sim::BlackBox)

Return custom metric of distance to event. For best results, metric should depend only on current state. For episodic simulation, return minimum distance to event across entire episode (miss distance).

distance(sim::GrayBox)

Return custom metric of distance to event. For best results, metric should depend only on current state. For episodic simulation, return minimum distance to event across entire episode (miss distance).

environment(sim::GrayBox)

Return Environment object constructed by simulation. Environment is an alias of Dict{Symbol, Sampleable}.

Step simulation and return log probability of transition.

flatten(distribution, value)

Flatten environment variable to quasi-normalized array.

Flatten EnvironmentValue into array.

generate_mdp(client::ASTClient)

Query remote MDP and generate RemoteASTMDP with matching attributes to interface with appropriate solvers.

Infer information about simulation environment.

Infer type of state.

isevent(sim::BlackBox)

Check whether simulation is in event (failure) state. For episodic simulation, check whether event occurred during episode.

isevent(sim::GrayBox)

Check whether simulation is in event (failure) state. For episodic simulation, check whether event occurred during episode.

isterminal(sim::BlackBox)

Check whether simulation has finished due to time limit or terminal state, independent of failure status.

isterminal(sim::GrayBox)

Check whether simulation has finished due to time limit or terminal state, independent of failure status.

Calculate log probability of sample from environment variable.

Calculate total log probability of environment value.

Infer dimension of state space.

Return observation of environment.

observe(sim::BlackBox)

Return quasi-normalized observation of simulation. Observation should copy, not reference, simulation state. If simulation is unobservable, leave unimplemented.

observe(sim::GrayBox)

Return quasi-normalized observation of simulation. Observation should copy, not reference, simulation state. If simulation is unobservable, leave unimplemented.

Open SSH tunnel to/from remote server.

Return ordered list dictionary keys.

ping(client::ASTClient)

Request empty echo from ASTServer, returning approximate round-trip time in seconds.

Salt random seed with private token for additional security.

Reset MDP.

reset!(sim::BlackBox)

Reset simulation.

reset!(sim::GrayBox)

Reset simulation.

Process simulation request from client and construct response.

reward(sim::BlackBox)

Advanced option: additional reward or heuristic. Returns scalar value or partially-applied function, allowing efficient calculation of R(s,s') = R(s)(s'). Should not alter simulation. See documentation for more details.

reward(sim::GrayBox, x::EnvironmentValue)

Advanced option: additional reward or heuristic. Returns scalar value or partially-applied function, allowing efficient calculation of R(s,a,s') = R(s,a)(s'). Should not alter simulation. See documentation for more details.

Listen for incoming requests and execute function calls on MDP. Can handle multiple client connections asynchronously, but server currently holds a single MDP instance.

set_password(server::ASTServer)

Set private token to string input (password). Avoids explicit representation of password in storage or side effects.

set_token(server::ASTServer, token)

Set private token to encrypt seeds. Recommended to use set_password instead, to avoid leaking token through side effects.

Return state type.

step([rng::AbstractRNG], sim::BlackBox)

Step simulation and return log probability of transition. If simulation uses global RNG, the one-argument step! should be implemented. If simulation requires non-global RNG, the two-argument function should be implemented instead.

step(sim::GrayBox, x::EnvironmentValue)

Step simulation given an EnvironmentValue object. EnvironmentValue is an alias of Dict{Symbol, Any}.

Read from socket until communication is synchronized. Typically unnecessary, but mitigates issue in which IO between client and server can become mismatched.

Return Boolean indicating termination status.

Reconstruct EnvironmentValue from array.

unflatten(distribution, array::AbstractVector{<:Real})

Reconstruct environment variable from quasi-normalized array.

Sample environment, returning EnvironmentValue (default) or array.

Analysis framework. Provides utilities to analyze and visualize solver results.

Simple uninformed Monte Carlo search. Should be used as a baseline for other local solvers, in terms of runtime performance and overall efficacy.

MCS <: LocalSolver

Monte Carlo search solver.

MCSResult <: LocalResult

Monte Carlo search result.

Solvers module. Provides algorithms for solving AST MDPs.

Abstract type for global result (adversarial policy).

Abstract type for solvers that return an adversarial policy.

Abstract type for local result (failure examples).

Abstract type for solvers that return failure examples.

Abstract type for solver output.

Abstract base type for solvers.

solve(::Solver, ::Function)

Apply solver to MDP generator function.

Delete least significant keys if length exceeds k. Facilitates top-k priority queueing.

Monte Carlo tree search algorithm. Balances exploration and experience. Constructs partially-ordered search tree over actions, treating states as anonymous.

MCTS <: LocalSolver

Monte Carlo tree search algorithm.

MCTSResult <: LocalResult

Monte Carlo tree search result.

Anonymous state node defined by unique chain of preceding actions. Search tree is composed of singly- and doubly-linked nodes.

Take an action from a given state, adding and returning new child. For transient action chains, such as rollouts, set forward link to false; nodes are deleted automatically by garbage collector when leaf goes out of scope.

Return unordered collection of children. Uses PQ backend to avoid unnecessary sort.

Return whether node is a leaf (no children).

Return whether node is a root (no parent).

Return maximum depth of tree.

Roll out random policy from given state, returning total return and final state.

Perform one full simulation of MCTS.

Return top-scoring action and next state.

Return total number of nodes in tree.

Return action chain from root to current node.